[PATCH v3 4/4] drm/xe: Add WA BB to capture active context utilization
Matthew Brost
matthew.brost at intel.com
Tue May 6 02:05:23 UTC 2025
On Mon, May 05, 2025 at 03:58:56PM -0700, Umesh Nerlige Ramappa wrote:
> Context Timestamp (CTX_TIMESTAMP) in the LRC accumulates the run ticks
> of the context, but only gets updated when the context switches out. In
> order to check how long a context has been active before it switches
> out, two things are required:
>
> (1) Determine if the context is running:
>
> To do so, we program the WA BB to set an initial value for CTX_TIMESTAMP
> in the LRC. The value chosen is 1 since 0 is the initial value when the
> LRC is initialized. During a query, we just check for this value to
> determine if the context is active. If the context switched out, it
> would overwrite this location with the actual CTX_TIMESTAMP MMIO value.
> Note that WA BB runs as the last part of the context restore, so reusing
> this LRC location will not clobber anything.
>
> (2) Calculate the time that the context has been active for:
>
> The CTX_TIMESTAMP ticks only when the context is active. If a context is
> active, we just use the CTX_TIMESTAMP MMIO as the new value of
> utilization. While doing so, we need to read the CTX_TIMESTAMP MMIO
> for the specific engine instance. Since we do not know which instance
> the context is running on until it is scheduled, we also read the
> ENGINE_ID MMIO in the WA BB and store it in the PPHSWP.
>
> Using the above 2 instructions in a WA BB, capture active context
> utilization.
>
> v2: (Matt Brost)
> - This breaks TDR, fix it by saving the CTX_TIMESTAMP register
> "drm/xe: Save CTX_TIMESTAMP mmio value instead of LRC value"
> - Drop tile from LRC if using gt
> "drm/xe: Save the gt pointer in LRC and drop the tile"
>
> v3:
> - Remove helpers for bb_per_ctx_ptr (Matt)
> - Add define for context active value (Matt)
> - Use 64 bit CTX TIMESTAMP for platforms that support it. For platforms
> that don't, live with the rare race. (Matt, Lucas)
> - Convert engine id to hwe and get the MMIO value (Lucas)
> - Correct commit message on when WA BB runs (Lucas)
>
> Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4532
> Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
> ---
> drivers/gpu/drm/xe/regs/xe_engine_regs.h | 4 +
> drivers/gpu/drm/xe/regs/xe_lrc_layout.h | 1 +
> drivers/gpu/drm/xe/xe_exec_queue.c | 2 +-
> drivers/gpu/drm/xe/xe_lrc.c | 153 ++++++++++++++++++++++-
> drivers/gpu/drm/xe/xe_lrc.h | 2 +-
> drivers/gpu/drm/xe/xe_lrc_types.h | 5 +-
> drivers/gpu/drm/xe/xe_trace_lrc.h | 8 +-
> 7 files changed, 163 insertions(+), 12 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/regs/xe_engine_regs.h b/drivers/gpu/drm/xe/regs/xe_engine_regs.h
> index 52f4c96c01dc..7ade41e2b7b3 100644
> --- a/drivers/gpu/drm/xe/regs/xe_engine_regs.h
> +++ b/drivers/gpu/drm/xe/regs/xe_engine_regs.h
> @@ -43,6 +43,10 @@
> #define XEHPC_BCS8_RING_BASE 0x3ee000
> #define GSCCS_RING_BASE 0x11a000
>
> +#define ENGINE_ID(base) XE_REG((base) + 0x8c)
> +#define ENGINE_INSTANCE_ID REG_GENMASK(9, 4)
> +#define ENGINE_CLASS_ID REG_GENMASK(2, 0)
> +
> #define RING_TAIL(base) XE_REG((base) + 0x30)
> #define TAIL_ADDR REG_GENMASK(20, 3)
>
> diff --git a/drivers/gpu/drm/xe/regs/xe_lrc_layout.h b/drivers/gpu/drm/xe/regs/xe_lrc_layout.h
> index 8619244b7c7b..994af591a2e8 100644
> --- a/drivers/gpu/drm/xe/regs/xe_lrc_layout.h
> +++ b/drivers/gpu/drm/xe/regs/xe_lrc_layout.h
> @@ -11,6 +11,7 @@
> #define CTX_RING_TAIL (0x06 + 1)
> #define CTX_RING_START (0x08 + 1)
> #define CTX_RING_CTL (0x0a + 1)
> +#define CTX_BB_PER_CTX_PTR (0x12 + 1)
> #define CTX_TIMESTAMP (0x22 + 1)
> #define CTX_TIMESTAMP_UDW (0x24 + 1)
> #define CTX_INDIRECT_RING_STATE (0x26 + 1)
> diff --git a/drivers/gpu/drm/xe/xe_exec_queue.c b/drivers/gpu/drm/xe/xe_exec_queue.c
> index 606922d9dd73..cd9b1c32f30f 100644
> --- a/drivers/gpu/drm/xe/xe_exec_queue.c
> +++ b/drivers/gpu/drm/xe/xe_exec_queue.c
> @@ -830,7 +830,7 @@ void xe_exec_queue_update_run_ticks(struct xe_exec_queue *q)
> {
> struct xe_device *xe = gt_to_xe(q->gt);
> struct xe_lrc *lrc;
> - u32 old_ts, new_ts;
> + u64 old_ts, new_ts;
> int idx;
>
> /*
> diff --git a/drivers/gpu/drm/xe/xe_lrc.c b/drivers/gpu/drm/xe/xe_lrc.c
> index ae6926dc900e..7b5d5310fd27 100644
> --- a/drivers/gpu/drm/xe/xe_lrc.c
> +++ b/drivers/gpu/drm/xe/xe_lrc.c
> @@ -24,6 +24,7 @@
> #include "xe_hw_fence.h"
> #include "xe_map.h"
> #include "xe_memirq.h"
> +#include "xe_mmio.h"
> #include "xe_sriov.h"
> #include "xe_trace_lrc.h"
> #include "xe_vm.h"
> @@ -655,6 +656,7 @@ u32 xe_lrc_pphwsp_offset(struct xe_lrc *lrc)
> #define LRC_CTX_JOB_TIMESTAMP_OFFSET (LRC_START_SEQNO_PPHWSP_OFFSET + 8)
> #define LRC_CTX_JOB_TIMESTAMP_UDW_OFFSET (LRC_CTX_JOB_TIMESTAMP_OFFSET + 4)
> #define LRC_PARALLEL_PPHWSP_OFFSET 2048
> +#define LRC_ENGINE_ID_PPHWSP_OFFSET 2096
>
> u32 xe_lrc_regs_offset(struct xe_lrc *lrc)
> {
> @@ -704,6 +706,11 @@ static inline u32 __xe_lrc_parallel_offset(struct xe_lrc *lrc)
> return xe_lrc_pphwsp_offset(lrc) + LRC_PARALLEL_PPHWSP_OFFSET;
> }
>
> +static inline u32 __xe_lrc_engine_id_offset(struct xe_lrc *lrc)
> +{
> + return xe_lrc_pphwsp_offset(lrc) + LRC_ENGINE_ID_PPHWSP_OFFSET;
> +}
> +
> static u32 __xe_lrc_ctx_timestamp_offset(struct xe_lrc *lrc)
> {
> return __xe_lrc_regs_offset(lrc) + CTX_TIMESTAMP * sizeof(u32);
> @@ -745,6 +752,7 @@ DECL_MAP_ADDR_HELPERS(ctx_timestamp)
> DECL_MAP_ADDR_HELPERS(ctx_timestamp_udw)
> DECL_MAP_ADDR_HELPERS(parallel)
> DECL_MAP_ADDR_HELPERS(indirect_ring)
> +DECL_MAP_ADDR_HELPERS(engine_id)
>
> #undef DECL_MAP_ADDR_HELPERS
>
> @@ -932,6 +940,65 @@ static void xe_lrc_finish(struct xe_lrc *lrc)
> xe_bo_unpin(lrc->bo);
> xe_bo_unlock(lrc->bo);
> xe_bo_put(lrc->bo);
> + xe_bo_unpin_map_no_vm(lrc->bb_per_ctx_bo);
> +}
> +
> +/*
> + * xe_lrc_setup_utilization() - Setup wa bb to assist in calculating active
> + * context run ticks.
> + * @lrc: Pointer to the lrc.
> + *
> + * Context Timestamp (CTX_TIMESTAMP) in the lrc accumulates the run ticks of the
> + * context, but only gets updated when the context switches out. In order to
> + * check how long a context has been running before it switches out, two things
> + * are required:
> + *
> + * (1) Determine if the context is running
> + * To do so, we program the wa bb to set an initial value for CTX_TIMESTAMP in
> + * the lrc. The value chosen is 1 since 0 is the initial value when the lrc is
> + * initialized. During a query, we just check for this value to determine if the
> + * context is active. If the context switched out, it would overwrite this
> + * location with the actual CTX_TIMESTAMP MMIO value. Note that wa bb will run
> + * after the context has been restored, so reusing this lrc location will not
> + * clobber anything.
> + *
> + * (2) Calculate the time that the context has been running for
> + * The CTX_TIMESTAMP ticks only when the context is active. If a context is
> + * active, we just use the CTX_TIMESTAMP MMIO as the new value of utilization.
> + * While doing so, we need to read the CTX_TIMESTAMP MMIO from the specific
> + * engine instance. Since we do not know which instance the context is running
> + * on until it is scheduled, we also read the ENGINE_ID MMIO in the wa bb and
> + * store it in the PPHSWP.
> + */
> +#define CONTEXT_ACTIVE 1ULL
> +static void xe_lrc_setup_utilization(struct xe_lrc *lrc)
> +{
> + u32 *cmd;
> +
> + cmd = lrc->bb_per_ctx_bo->vmap.vaddr;
> +
> + *cmd++ = MI_STORE_REGISTER_MEM | MI_SRM_USE_GGTT | MI_SRM_ADD_CS_OFFSET;
> + *cmd++ = ENGINE_ID(0).addr;
> + *cmd++ = __xe_lrc_engine_id_ggtt_addr(lrc);
> + *cmd++ = 0;
> +
> + *cmd++ = MI_STORE_DATA_IMM | MI_SDI_GGTT | MI_SDI_NUM_DW(1);
> + *cmd++ = __xe_lrc_ctx_timestamp_ggtt_addr(lrc);
> + *cmd++ = 0;
> + *cmd++ = lower_32_bits(CONTEXT_ACTIVE);
> +
> + if (GRAPHICS_VER(lrc_to_xe(lrc)) >= 20) {
Same suggestion as the previous patch:
s/GRAPHICS_VER(...)/xe->info.has_64bit_ctx_timestamp ?
I think everything else LGTM.
Matt
> + *cmd++ = MI_STORE_DATA_IMM | MI_SDI_GGTT | MI_SDI_NUM_DW(1);
> + *cmd++ = __xe_lrc_ctx_timestamp_udw_ggtt_addr(lrc);
> + *cmd++ = 0;
> + *cmd++ = upper_32_bits(CONTEXT_ACTIVE);
> + }
> +
> + *cmd++ = MI_BATCH_BUFFER_END;
> +
> + xe_lrc_write_ctx_reg(lrc, CTX_BB_PER_CTX_PTR,
> + xe_bo_ggtt_addr(lrc->bb_per_ctx_bo) | 1);
> +
> }
>
> #define PVC_CTX_ASID (0x2e + 1)
> @@ -973,10 +1040,17 @@ static int xe_lrc_init(struct xe_lrc *lrc, struct xe_hw_engine *hwe,
> if (IS_ERR(lrc->bo))
> return PTR_ERR(lrc->bo);
>
> + lrc->bb_per_ctx_bo = xe_bo_create_pin_map(xe, tile, NULL, SZ_4K,
> + ttm_bo_type_kernel,
> + bo_flags);
> + if (IS_ERR(lrc->bb_per_ctx_bo)) {
> + err = PTR_ERR(lrc->bb_per_ctx_bo);
> + goto err_lrc_finish;
> + }
> +
> lrc->size = lrc_size;
> lrc->ring.size = ring_size;
> lrc->ring.tail = 0;
> - lrc->ctx_timestamp = 0;
>
> xe_hw_fence_ctx_init(&lrc->fence_ctx, hwe->gt,
> hwe->fence_irq, hwe->name);
> @@ -1049,7 +1123,13 @@ static int xe_lrc_init(struct xe_lrc *lrc, struct xe_hw_engine *hwe,
> xe_lrc_read_ctx_reg(lrc, CTX_CONTEXT_CONTROL) |
> _MASKED_BIT_ENABLE(CTX_CTRL_PXP_ENABLE));
>
> - xe_lrc_write_ctx_reg(lrc, CTX_TIMESTAMP, 0);
> + lrc->ctx_timestamp = CONTEXT_ACTIVE;
> + xe_lrc_write_ctx_reg(lrc, CTX_TIMESTAMP,
> + lower_32_bits(CONTEXT_ACTIVE));
> + if (GRAPHICS_VER(lrc_to_xe(lrc)) >= 20) {
> + xe_lrc_write_ctx_reg(lrc, CTX_TIMESTAMP_UDW,
> + upper_32_bits(CONTEXT_ACTIVE));
> + }
>
> if (xe->info.has_asid && vm)
> xe_lrc_write_ctx_reg(lrc, PVC_CTX_ASID, vm->usm.asid);
> @@ -1078,6 +1158,8 @@ static int xe_lrc_init(struct xe_lrc *lrc, struct xe_hw_engine *hwe,
> map = __xe_lrc_start_seqno_map(lrc);
> xe_map_write32(lrc_to_xe(lrc), &map, lrc->fence_ctx.next_seqno - 1);
>
> + xe_lrc_setup_utilization(lrc);
> +
> return 0;
>
> err_lrc_finish:
> @@ -1297,6 +1379,21 @@ struct iosys_map xe_lrc_parallel_map(struct xe_lrc *lrc)
> return __xe_lrc_parallel_map(lrc);
> }
>
> +/**
> + * xe_lrc_engine_id() - Read engine id value
> + * @lrc: Pointer to the lrc.
> + *
> + * Returns: context id value
> + */
> +static u32 xe_lrc_engine_id(struct xe_lrc *lrc)
> +{
> + struct xe_device *xe = lrc_to_xe(lrc);
> + struct iosys_map map;
> +
> + map = __xe_lrc_engine_id_map(lrc);
> + return xe_map_read32(xe, &map);
> +}
> +
> static int instr_dw(u32 cmd_header)
> {
> /* GFXPIPE "SINGLE_DW" opcodes are a single dword */
> @@ -1844,21 +1941,67 @@ void xe_lrc_snapshot_free(struct xe_lrc_snapshot *snapshot)
> kfree(snapshot);
> }
>
> +static int get_ctx_timestamp(struct xe_lrc *lrc, u32 engine_id, u64 *reg_ctx_ts)
> +{
> + u16 class = REG_FIELD_GET(ENGINE_CLASS_ID, engine_id);
> + u16 instance = REG_FIELD_GET(ENGINE_INSTANCE_ID, engine_id);
> + struct xe_hw_engine *hwe;
> + u64 val;
> +
> + hwe = xe_gt_hw_engine(lrc->gt, class, instance, false);
> + if (xe_gt_WARN_ONCE(lrc->gt, !hwe || xe_hw_engine_is_reserved(hwe),
> + "Unexpected engine class:instance %d:%d for context utilization\n",
> + class, instance))
> + return -1;
> +
> + if (GRAPHICS_VER(lrc_to_xe(lrc)) >= 20)
> + val = xe_mmio_read64_2x32(&hwe->gt->mmio,
> + RING_CTX_TIMESTAMP(hwe->mmio_base));
> + else
> + val = xe_mmio_read32(&hwe->gt->mmio,
> + RING_CTX_TIMESTAMP(hwe->mmio_base));
> +
> + *reg_ctx_ts = val;
> +
> + return 0;
> +}
> +
> /**
> * xe_lrc_update_timestamp() - Update ctx timestamp
> * @lrc: Pointer to the lrc.
> * @old_ts: Old timestamp value
> *
> * Populate @old_ts current saved ctx timestamp, read new ctx timestamp and
> - * update saved value.
> + * update saved value. With support for active contexts, the calculation may be
> + * slightly racy, so follow a read-again logic to ensure that the context is
> + * still active before returning the right timestamp.
> *
> * Returns: New ctx timestamp value
> */
> -u32 xe_lrc_update_timestamp(struct xe_lrc *lrc, u32 *old_ts)
> +u64 xe_lrc_update_timestamp(struct xe_lrc *lrc, u64 *old_ts)
> {
> + u64 lrc_ts, reg_ts;
> + u32 engine_id;
> +
> *old_ts = lrc->ctx_timestamp;
>
> - lrc->ctx_timestamp = xe_lrc_ctx_timestamp(lrc);
> + lrc_ts = xe_lrc_ctx_timestamp(lrc);
> + if (lrc_ts == CONTEXT_ACTIVE) {
> + engine_id = xe_lrc_engine_id(lrc);
> + if (!get_ctx_timestamp(lrc, engine_id, ®_ts))
> + lrc->ctx_timestamp = reg_ts;
> +
> + /* read lrc again to ensure context is still active */
> + lrc_ts = xe_lrc_ctx_timestamp(lrc);
> + }
> +
> + /*
> + * If context switched out, just use the lrc_ts. Note that this needs to
> + * be a separate if condition. Do not convert this to an else block
> + * corresponding to the above if.
> + */
> + if (lrc_ts != CONTEXT_ACTIVE)
> + lrc->ctx_timestamp = lrc_ts;
>
> trace_xe_lrc_update_timestamp(lrc, *old_ts);
>
> diff --git a/drivers/gpu/drm/xe/xe_lrc.h b/drivers/gpu/drm/xe/xe_lrc.h
> index 513093c2b5d6..a0661ac6b1bc 100644
> --- a/drivers/gpu/drm/xe/xe_lrc.h
> +++ b/drivers/gpu/drm/xe/xe_lrc.h
> @@ -138,6 +138,6 @@ u64 xe_lrc_ctx_job_timestamp(struct xe_lrc *lrc);
> *
> * Returns the current LRC timestamp
> */
> -u32 xe_lrc_update_timestamp(struct xe_lrc *lrc, u32 *old_ts);
> +u64 xe_lrc_update_timestamp(struct xe_lrc *lrc, u64 *old_ts);
>
> #endif
> diff --git a/drivers/gpu/drm/xe/xe_lrc_types.h b/drivers/gpu/drm/xe/xe_lrc_types.h
> index cd38586ae989..ae24cf6f8dd9 100644
> --- a/drivers/gpu/drm/xe/xe_lrc_types.h
> +++ b/drivers/gpu/drm/xe/xe_lrc_types.h
> @@ -52,7 +52,10 @@ struct xe_lrc {
> struct xe_hw_fence_ctx fence_ctx;
>
> /** @ctx_timestamp: readout value of CTX_TIMESTAMP on last update */
> - u32 ctx_timestamp;
> + u64 ctx_timestamp;
> +
> + /** @bb_per_ctx_bo: buffer object for per context batch wa buffer */
> + struct xe_bo *bb_per_ctx_bo;
> };
>
> struct xe_lrc_snapshot;
> diff --git a/drivers/gpu/drm/xe/xe_trace_lrc.h b/drivers/gpu/drm/xe/xe_trace_lrc.h
> index 5c669a0b2180..d525cbee1e34 100644
> --- a/drivers/gpu/drm/xe/xe_trace_lrc.h
> +++ b/drivers/gpu/drm/xe/xe_trace_lrc.h
> @@ -19,12 +19,12 @@
> #define __dev_name_lrc(lrc) dev_name(gt_to_xe((lrc)->fence_ctx.gt)->drm.dev)
>
> TRACE_EVENT(xe_lrc_update_timestamp,
> - TP_PROTO(struct xe_lrc *lrc, uint32_t old),
> + TP_PROTO(struct xe_lrc *lrc, uint64_t old),
> TP_ARGS(lrc, old),
> TP_STRUCT__entry(
> __field(struct xe_lrc *, lrc)
> - __field(u32, old)
> - __field(u32, new)
> + __field(u64, old)
> + __field(u64, new)
> __string(name, lrc->fence_ctx.name)
> __string(device_id, __dev_name_lrc(lrc))
> ),
> @@ -36,7 +36,7 @@ TRACE_EVENT(xe_lrc_update_timestamp,
> __assign_str(name);
> __assign_str(device_id);
> ),
> - TP_printk("lrc=:%p lrc->name=%s old=%u new=%u device_id:%s",
> + TP_printk("lrc=:%p lrc->name=%s old=%llu new=%llu device_id:%s",
> __entry->lrc, __get_str(name),
> __entry->old, __entry->new,
> __get_str(device_id))
> --
> 2.43.0
>
More information about the Intel-xe
mailing list